Summary

庄明浩 argues that the “AI games” worth discussing are not games with AI inserted into art, code, or marketing pipelines, but whether generative AI can change gameplay. Game AI has existed for decades; by September 2024, Google Cloud had even organized the capabilities across the stack into a standard matrix, while 37Games, 恺英网络 and others repeatedly highlighted their AI middle platforms in financial reports. “If this is what people were expecting, then it’s already over.”
AI games will be hard to establish as a clearly bounded new category in the near term; a category explicitly called “AI games” may not emerge at all. The core problem is that gameplay itself cannot be classified consistently: anime-style labels describe world-building and art, extraction shooters describe mechanics, and UGC describes content origin and platform attributes. “We narrowed the question down to gameplay, only to find that we can’t define gameplay.”
Historical gameplay innovation is usually a recombination of old labels plus a small innovation, and AI has no natural claim to that role. Shooters evolved into battle royale, then added resource competition to become extraction shooters; MOBA itself evolved from RTS. AI is being put at the center largely because it is “easier to finance or easier to attract attention,” not because it is necessarily the answer.
Even the richest and most risk-tolerant publishers have so far delivered only localized changes such as dialogue, NPCs, and UGC generation. miHoYo’s dialogue-puzzle test product turns most of the experience into a ChatGPT-like chat window; its UE5 open-world trailer planted the line “Attention Is All You Need,” though it remains unclear whether that was merely a joke or something deeper. Yuanxiang has likewise moved from supplying AI NPC technology to developing 《昭阳传》, without yet revealing its end state.
World models have reignited end-state visions of Ready Player One and Westworld, while the adjacent path of AI companionship has already been judged by 庄明浩 as “basically a total rout.” “There should be 12” of a16z’s top 50 external applications were social-companion products, but sustained engagement, retention, and monetization remain difficult outside borderline content. A former Cocos executive’s old line offers a brutal test: “A game has a win condition; everything without one is called the metaverse.”
The biggest problem with this crop of projects may not be the models, but founders’ “lack of awe for the act of making games.” Games are mature products worth several hundred billion dollars globally and roughly RMB300B in China, while also delivering increasingly service-like experiences; leading titles take hundreds or even thousands of people years to build. Saying “we’ll sell skins and stories” does not mean understanding game monetization. “Can you really just learn how to make 暖暖?”
With short-term narratives that cannot be falsified, the venture market can only retreat to founder track records and person-role fit, with fear of missing out driving successive rounds of financing. A 10-to-20-person team costing roughly RMB5M a year and not training a base model can raise more than $20M; at 4% interest plus fixed-income returns, “the interest in the account is enough to support my team.” That removes urgency from product validation, while the few projects that avoid dying in public continue to attract bids because supply is so scarce.

Deep dive

1. This Wave of AI Games Must First Redraw the Boundary Around Generative AI

庄明浩 first pushed back on compressing AI’s history into the 3 years since ChatGPT appeared. The technology moved from rule-based systems and machine learning through neural networks to Transformers, GPT-3, and ChatGPT based on GPT-3.5; AI history itself can be traced to roughly the 1940s or 1950s. Under his framing, this round of discussion may focus on how generative AI centered on language and multimodality combines with games.
Game AI has also been around for a long time. In Pac-Man, the 4 enemies with different colors adopt different strategies when the player comes within 3, 6, or 1 tiles; in FIFA, teammate positioning, opponent pressure, shots, and slide tackles are all AI in the broad sense, not capabilities that appeared only today.
Natural-language models are unquestionably at the center of this round’s discussion. After ChatGPT appeared, one expert in the industry said, “Once machines master language, artificial general intelligence has arrived.” AI coding can be viewed as a subset of language, or as another core language for building virtual worlds. 庄明浩 relayed the Anthropic founder’s view that within 12 months, AI might reach a point where it can write all code.
Multimodality is now converging from the separate battlefields of images, video, audio, and 3D into world models. After 李飞飞’s startup released its first test product, a former classmate with a fear of heights asked, “Once I use your product to build that world, can I play in it?” She replied that humanity may really be building that “real, silicon-based world”—with “real” in quotation marks.

2. AI Production Tools Had Become Infrastructure by 2024

The most direct combination is AI entering game production: generating 2D and 3D assets, art, characters, code, music, and video, then extending into marketing and player experience. The overwhelming majority of companies calling themselves AI game companies are still operating somewhere along this production-tool chain.
In September 2024, Google Cloud organized these capabilities into a four-quadrant framework spanning production, marketing, player experience, and cost and difficulty. What mattered to 庄明浩 was not the chart’s contents but the timing: “When something can already be summarized and categorized like this, it means it’s no longer new.”
Over the following 1 to 2 years, “AI middle platforms” appeared repeatedly in the financial reports of 37Games, 恺英网络 and others. Virtually every game company of meaningful scale had adopted them, while AI could already handle essentially all environment art in 2D-heavy settings. This can cut costs and improve efficiency, but it is not a gameplay thesis for an AI-native game.

3. Gameplay Itself Cannot Be Clearly Classified, Making AI Games Harder to Define as a Category

Games now combine extreme industrialization with extreme creativity: the global market is worth several hundred billion dollars, while China’s is roughly RMB300B. Leading projects take hundreds or even thousands of people years to complete; small teams have to seek extreme creativity within genres that have already been tried repeatedly. Games today are closer to products, while also providing increasingly service-like experiences.
The China Audio and Digital Association’s standards classify games by subject matter, gameplay, source, technology, form of presentation, creative nature, and player scale. Gameplay itself is divided into 16 major categories, including role-playing, competitive, strategy, cards, board and casual games, tower defense, and match-three. Yet many products fit into several boxes at once; labels are not the same thing as gameplay.
Anime-style games describe world-building and art more than mechanics. Extraction shooters are more clearly a gameplay label, but do not appear in the classification above and may also contain other elements. Battle royale likewise cannot capture the full experience. UGC is closer to a description of platform attributes and content origin. 庄明浩’s paradox is: “We want AI to combine with gameplay, but what is gameplay? We can’t define it.”
He therefore believes it will be difficult to draw a clear boundary around an AI game category in the near term. Even if a label called “AI games” emerges in the future, it is unlikely to become a clear, stable classification anytime soon. This is not a claim that AI will not enter games, but that technology origin alone is unlikely to become a gameplay label.

4. New Gameplay Comes From Recombining Old Labels; AI Is Merely a Candidate for the Micro-Innovation

庄明浩 summarizes game evolution as “a recombination of old labels with a little micro-innovation.” The progression from shooters to battle royale to extraction shooters, and the evolution of MOBA from RTS, show that commercial games more often rely on continuous innovation than on entirely new genres appearing from nowhere.
On that logic, an AI game might combine one or more old labels—sandbox, Roguelike, editor, tabletop role-playing, puzzle, text adventure, or Werewolf—with AI providing the micro-innovation. These labels are repeatedly selected because the industry broadly assumes they are easier to combine with AI. But his more important question is: “Why does AI have to take on the micro-innovation? Why can’t something else do it?”
Most AI game coverage today can be mapped back onto these old-label combinations. The smallest step is putting AI into NPCs or helping players generate 3D assets within a UGC mode. 庄明浩 uses Eggy Party as an example: because generation happens on the player consumption and creation side rather than inside the production middle platform, it does touch gameplay—but still amounts to “micro-micro-micro-micro-micro innovation.”

5. Even the Richest Publishers Have Delivered Only Micro-Innovations

Innovation requires the most money, people, and capacity to absorb risk, so 庄明浩 looks at miHoYo and Yuanxiang as companies that can afford to be “任性.” The former has an almost unlimited budget, its founder as project lead, and no monetization or listing pressure; the latter still held several hundred million dollars in cash after its previous metaverse financing round.
miHoYo’s product tested on Steam is a dialogue-puzzle game: players converse with NPCs to find a way off a planet, and aside from some scenes, the game’s visual experience is essentially a ChatGPT-like chat window. 庄明浩 leaves room for judgment: “He may be doing the right thing,” but the product remains far from the innovation people imagine.
The newly announced UE5 open-world project includes roughly 30 minutes of gameplay footage. A warning strip in the scene reads, “Attention Is All You Need.” 庄明浩 is unsure whether this is merely a joke or whether the team has actually hidden something deeper.
Yuanxiang was founded around 2019 to 2020, with a founder who had been head of Tencent AI Lab. Before ChatGPT became a phenomenon, it entered the metaverse through AI NPCs and raised several hundred million dollars. After the bubble receded and real AI arrived, it moved one step beyond being a technology provider and began making the conversational alternate-history game 《昭阳传》 itself.

6. World Models Restart the End-State Narrative, While AI Companionship Is Basically a Total Rout

World models have brought end-state visions from Ready Player One, Westworld, and Sword Art Online back into the conversation, but reality remains at a very early stage. 庄明浩 cites a capability test that assigned an AI Agent the role of customer service for a DIY computer shop: even the strongest model had acquired some common sense, but remained “far, far, far” from human-level common-sense cognition.
AI social products and companions are viewed as another route toward the same end state. “There should be 12” of a16z’s top 50 external AI applications belonged to this category. Character.AI, Talkie, 星野, and Tolan, which was discussed in the first half of this year, have all represented moments of market excitement.
Tolan is reportedly a small-alien companion product from a French company. China has also seen imitations and similar products, including EVA and other AI companion offerings. Domestic teams such as Kunlun, 作业帮, Yuanxiang, and 西湖心辰 have tried similar approaches. 庄明浩 adds an important qualifier—“many companies are still trying”—but as of now, “it’s basically a total rout.” Outside borderline interactions, these products appear difficult to sustain for long, and it remains hard to prove that retention and monetization can support platform operations.
He closes the distinction with a definition from the Cocos founder during the metaverse boom: “Things produced with our game engine that have a win condition are games; everything without a win condition is collectively called the metaverse.” In his view, the failures of many later projects reinforced the criticism that their founders “lack awe for the act of making games.”

7. The Falsifiability Vacuum Pushes the Venture Market Back to “People Only”

If these projects cannot be logically falsified in the near term, then “whatever the founder says is right.” Investors still need to answer whether retention can support a platform, whether existing monetization methods can connect to the product, and whether the project truly should not have path dependence.
吴秉坚 of New Capital Partners summarized the landscape on January 13, 2023: consumer startups built on large models fall into 3 types—ChatGPT variants, Character.AI variants, and Stanford Town variants. 庄明浩 believes this old 2023 slide still applies today.
Unable to evaluate the business, early-stage VCs retreat to founder backgrounds and the fit between past experience and the current direction. This is “a correct but useless truism,” yet it leaves accomplished founders with another curse: they are committed to an ultimate end state and can continue raising money on the strength of their backgrounds alone.
He cites several founders: 吴萌, former CEO of 即刻, is working on an AI 3D game engine or a generated game world; 郭列, former founder of 脸萌, made a game similar to Eggy Party but has since shut it down; and 王诗沐, founder of NetEase Cloud Music, built a 3D generative-AI product similar to Character.AI.
The result is a loop: investors fear missing out, projects raise round after round on the strength of their people, and then may “die in the light of day” once the product launches. Conversely, any project that survives several rounds of discussion, produces something, and does not completely die can attract a premium because supply is so limited. “The market is determined by both sides of supply; there is no right or wrong.”

8. Ample Funding Delays Validation, While Monetization Analogies Expose Gaps in Game Literacy

One founder put it plainly: “The interest on the money in my account is enough to support my team.” 庄明浩 runs the numbers: a 10-to-20-person team that is not training a base model needs mostly payroll plus a small amount of compute, with RMB5M enough for a year. If it raises more than $20M, then at 4% interest plus fixed-income returns, the interest alone covers the team. “So what’s the rush?”
One investor described the market this way: “Put AI into games, put games into AI, and then there are some other products that insist on calling themselves AI games.” That maps closely to the production-tool, gameplay-micro-innovation, companionship, and metaverse narratives 庄明浩 has separated out.
One project used DAU, ARPU, and other metrics from an existing MMO to model future monetization; his only comment was, “Simple.” Another team proposed “selling skins and selling stories,” prompting his response: “Can you just learn how to make 暖暖?” His point was that such monetization plans may underestimate the content and product capabilities required to make a game.
The final conclusion is not that any particular company is doomed, but that at this point “pure business discussion is meaningless.” The logic cannot yet be falsified, so investment decisions return to what the founder has done before and how much he is worth. Investors and founders both know this, but “when you’re in the arena, all you can say is that what exists is justified; there is no right or wrong.”

Claude Opus 4.5, Genesis Mission and Amazon's $50B AI Push ChatGPT Groupchats, System Prompts, and Daily LLM Use Cases | Sharp Tech with Ben Thompson